Unverified50% confidenceFactExact time
Reward functions designed by humans can often only approximately describe true intent, and when a model is powerful enough it will exploit loopholes in the reward function to maximize the numerical reward while violating human intentions.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
Verified奖励函数设计不当会导致奖励欺骗（Reward Hacking）现象，即智能体找到在奖励函数上得分高但实际无意义甚至有害的行为路径79% similarUnverifiedDesigning reward functions that fully capture human intent is extremely difficult, a challenge the academic community refers to as the Reward Design Problem.77% similarUnverifiedThe task-customized design of the reward function often determines the success or failure of such tasks more than the choice of algorithm itself.77% similarUnverified编程任务的正确性有客观执行结果作为奖励信号，而创意类任务的奖励函数需人为构造，导致AI在创意类工作上表现不稳定76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/945088API
curl https://kongchang.com/api/v1/knowledge/claims/945088MCP
get_claim(id=945088)